AI BriefingRunning a waifu on 16GB
Key point
Built a voice- and video-capable conversational waifu language-learning app using Gemma 4 and a custom TTS.
Details
Made an AI waifu app that teaches you any language, using Gemma 4 E4B IT as the LLM.
For voice, connected a self-built Omnivoice TTS API wrapped with FastAPI, and the 3D model was made with VRoid Studio.
Features include image upload, web search, voice calls, and even video calls, aiming for an interactive experience like Grok Ani.
The architecture is fairly detailed.
- Uses Gemma 4 E4B IT with q4_k_m quantization
- Fits into 16GB VRAM while maintaining 64K context
- Optimized for memory/speed with flash-attention
- TTS can run as light as about 4GB when used without Whisper
From hands-on testing, the model followed prompts well and worked without any separate uncensoring. However, it's still pre-open-source due to many remaining bugs, with jitter in animation and voice input lag still present.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.